Community LoRA Sharpens MiniMax H3 Videos Without Hosted 2K API
A community LoRA for MiniMax H3 sharpens video output through aligned guide latents, offering a new pass on ComfyUI's ref2va pipeline.
Takeaways
- Rank-64 LoRA for MiniMax H3 ref2va that sharpens video while keeping photorealism.
- Conditions via aligned guide latents through the native
MiniMaxH3AddGuidenode, not H3's reference-video block. - Guide frame i maps to target frame i at identical resolution, eliminating positional search cost in attention.
- Training kept the guide near-clean (0.1% noise) while noising the target normally, forcing a direct mapping.
- Requires H3's valid clip lengths (5, 22, 39, 56, 73, 90… frames); mismatched length or resolution breaks it.
- Apache 2.0, intended as a second-pass sharpener rather than a from-scratch generator.
A community-trained LoRA now turns MiniMax H3 into a source-to-source video sharpening tool. Alissonerdx’s ComfyUI adapter uses a rank-64 network to add crisp detail while preserving a photorealistic appearance. For developers running H3 locally, it offers a second-pass enhancement workflow that avoids MiniMax’s hosted 2K regeneration stage.
MiniMax H3 is a 33-billion-parameter multimodal model that generates video and stereo audio in a single pass. MiniMax released open weights with native ComfyUI support, split across two checkpoints: FL2VA handles text-to-video and first-and-last-frame conditioning, while Ref2VA supports reference-based generation. The LoRA was trained specifically for Ref2VA.
Aligned latents do the sharpening
The model card specifies a literal trigger caption: enhance a video with sharp, crisp details while preserving a natural photorealistic appearance. The adapter receives the source clip as guide latents, which are compressed video representations passed directly to H3’s transformer.